gameplay-ai

[fading shale]

it didn't work 😦

I'll get a screenshot...my UE is on an offline PC

[fading shale]

ok, so the gamemode tells the manager in the level to spawn and setup AI

the manager is an actor in the level that will handle spawning aliens when they need to be spawned.

and the aicontroller does what aicontrollers do.

you'll see I have two print strings after the move to location node...currently when the alien spawns, I get about 15 printouts of the XYZ location of the two different spawnpoints I have set up, randomly since it's picking randomly between the two right now.

so it's failing to move and telling me where it tried to move to.

with this setup...still, if I place a monster in the level at design time, when I run, it immediately starts moving and outputs Moved! when it reaches a nav target

[cinder garnet]

if anyone can help @ me, this enemy AI is supposed to follow you while you run away, he does all that until he gets to the door, then he will sit there and stare even with both sliding doors completely open

[uneven cloud]

If you hit F8 and turn on navigation, is there a nav mesh through the door?

[dense condor]

Does anyone know why spawned AI doesnt move?

I've tried anything i could find online. I have mesh invokers set up to my character, I have Generate Navigation only around nav invokers, I have my auto posses ai set to "Placed in world or spawned", i have spawn default controller on begin play, and nothing seems to work.

Placing ai in world works fine.

Also i have runtime generation and everything set up as dynamic as well.

[fading shale]

I'm having this same problem! see my code above.

[dense condor]

im usin behaviour trees and using the generic MoveTo task.

[fading shale]

I think I got it working...can you check in your movement controller the nav agent settings under movement capabilities? I had all -1 for radius and height...setting those to closer to real values made him move on spawn

[cinder garnet]

My world files just got corrupted 😢 so trying to see if I can get it back, any recommendations let me know

[fading shale]

but now it's not avoiding obstacles and goign right through them

[dense condor]

that didnt work either. It looks like my begin play is not working for some reason. I didnt notice before but if you spawn an AI character, is it supposed to spawn an ai controller in the world as well?

[fading shale]

Yes. And I actually stopped using begin play and created an interface event that does the ai setup and command to begin moving

[lyric flint]

You’ll want to make sure possess on spawn (can’t remember the exact name) is set on the character

[fading shale]

Anyone know how to make an aicontroller driven pawn with floating pawn movement dodge obstacles? I have a bunch of static meshes with collision and the ai pawn also has collision (I can hit it with my projectile) but it's navigation efforts just pass right through things in the scene

All of my scene static meshes are spawned from a routine not placed in the level at design time if that matters

[turbid reef]

Hi, anyone can help me with AI of a racing car, it needs modifications. I can pay. Please dm

[crimson oxide]

Hey there Im having a weird issue with behavior trees. So I have break the links from the root, saved, but the behavior tree is still playing as normal. Is this a compilere bug thing (?

[terse panther]

It works 😅, you need to disconnect the sequence/selector node right after the root .... I too once got stuck at this

[golden cove]

Hey there! Do we have an open source dataset for rigged 3d humans/metahumans?? (Please ping if you reply(

[lone stump]

it's a bug https://issues.unrealengine.com/issue/UE-97749

[ocean wren]

Hmm, so I got a chatgpt based metahuman working.. which was kind of fun

uses amazon polly to speak, uses deepgram to listen to my voice

ChatGPT api is ok funnily enough

But it annoys the crap out of me that they baked the "as an AI model" responses into the API

obviously I need to nail the initial prompt better

I also need to tell it to use shorter responses 🙂 it does like to bang on 😉

[celest python]

I know right

[ocean wren]

Oh my.. using playground to figure out my prompt settings, I asked it to speak like a nun 🙂

I need something.. sort of human-like, but sports commentator, nun and cyborg aren't cutting it, oh and gangster is a defite no-go

Think I've got it...

"You are a helpful friendly person called Polly and you have a friend called Phil who you are chatting to, please keep your responses short and speak like a seductress who really likes money"

🙂

hmm, not sure the "really likes money" thing worked..

Cutting video together now.. this stuff is kind of interesting

[ocean wren]

https://youtu.be/ja-8gfgsG0g

[celest python]

haha nice 😄 GPT vibes can be seen in responses though

[analog root]

after reaching high rounds, Some AIs don't move after spawning, everything works in the first rounds but in the rounds after 9 because of the numbers of AIs that becomes a lot, some AIs can not move, any idea?

they just staying in idle and not moving

[ocean wren]

Wait for it.. doing another attempt.. it matters what you use as the initial prompt 🙂

Testing this one next: "You are a disillusioned old hag called polly, everything is too much effort for you. You hate everyone, but you have a special hatred of Phil, who you now have the displeasure of talking to, respond in an evil tone, using as much contempt as you can muster."

And trust me, this delivers 🙂

[celest python]

check visual logger

[dense condor]

it ended up being a casting failure

[ocean wren]

https://youtu.be/oiyFTYQ07Vs

I lose it towards the end 🙂 I figure this is my life now.. finding stupid ways to coach AI into hating me 🙂

[scarlet sky]

Is it normal that the behavior tree blinks every flame like crazy?

[ocean wren]

No

[celest python]

some selector is going crazy

[scarlet sky]

Hmm good to know, mine is blinking a lot. How can I debug the BT?

tbh my bt is not even complex

zoomed a bit

above that there's only the root

[ocean wren]

it usually blinks because its failing to select something to run

[scarlet sky]

so it keeps on retrying?

[ocean wren]

yeah, selector will keep trying until it can run something

I'd usually have an idle that would be the lowest priority that I would fall into unless something higher takes over

[scarlet sky]

I guess it also can be heavy on performance the constant retry, right? As it keeps verifying conditions

[scarlet sky]

How did you made that? I'm a bit lost on how to escape this xD

[ocean wren]

Make an idle sequence.. that just plays an animation and then calls a wait node for a second or so

The point is that your other nodes will take over priority from the idle when they detect some condition that allows it.. see an enemy or whatever

otherwise, they go back to idle..

[scarlet sky]

I'll try to do that, still have to understand how priority works correctly here haha

[ocean wren]

left to right.. so right should have your "idle", then anything left of that takes over based on some condition being set.. seeing enemy, not being at home base etc..

you need to think of what your priority order is

and what conditions control which priorities are allowed

for instance, can't fire on an enemy if out of ammo, so getting ammo when out of ammo is probably higher priority than firing

that kind of thing

[scarlet sky]

Oh I see, so the EQS itself will never break the Idle even though it has a interval, because it does not has the priority condition like a blackboard decorator node

Yeah I have no clue on how to put a priority on that EQS, looked some decorators such as cooldown but it runs once between the cooldown, making movement impossible

I'll remove the interval on the EQS and make a timer on my BP itself, then set it as a float, probably terrible idea but I have no clue for now

[ocean wren]

The roam sequence could have a cooldown condition

then it would run at higher priority, unless it needed to cool down

in which case the idle would get a chance to run

The basic idea, is that you have conditions that determine which part of the tree executes

[scarlet sky]

Sorry, I don't quite get this part, with the cooldown condition wouldn't it always cooldown after running it? What you mean with "needed" to cooldown hm

[ocean wren]

cooldown is when something runs but you don't want it to run again immediately, so you cooldown (essentially tell the tree not to run this part) for a while until it is time to run again

that way, the idle would then be selected

and you'd get some idle for a while.. then the cooldown timer would run out and you could run the higher priority again

[uneven cloud]

Your problem here is that the move to starts before your EQS will finish. You need a task for the EQS first and then move to.

[scarlet sky]

But isn't the RunEQS there already... The EQS?

[uneven cloud]

The EQS doesn't return immediately. It's in a service, which happens while the task is running.

[scarlet sky]

I'm sorry I'm such a dead brain btw

I'll see some videos teaching about the system itself, I feel I don't quite understand

For example setting cooldown makes movement impossible because it needs to constantly be called

I think my whole setup is wrong hm

[olive prism]

How i can add a BTS in selector like this?

[scarlet sky]

But I appreciate you guys so so much

[olive prism]

Help

I'm falling into despair

[tiny comet]

I am just starting to use the new StateTree - Is there any notion of an event triggering a state change? I am trying to avoid polling state on tick

[analog root]

I use it but still don't know what is going on and why some of the AIs do not move to the player🫤

[analog root]

LogAIPerception (Error) Invalid sight targets found during UAISense_Sight::Update call
I see this error , any idea?

[vernal jacinth]

when using AI perception, how can I change the PeripheralVisionAngleDegrees at runtime?

epic's code comment says

    /** How far to the side AI can see, in degrees. Use SetPeripheralVisionAngle to change the value at runtime. 

but the SetPeripheralVisionAngle function does not exist, it only exists on the PawnSensingComponent, which I'm not using

[fading shale]

I have an aicontroller driving a floating pawn movement...it just randomly chooses targetpoints and moves to them. I've recently modified some settings trying to get it to move when spawned as opposed to being placed in the level, and now it no longer dodges other static meshes in my scene...the static meshes are spawned if that matters.

[stoic pier]

how is move to acceptable radius calculated, capsule to capsule?

or is it based on how you setup the collision presets?

I tried to add an extra 'pawn' sphere collision to the snout of this beast thinking it would offset the acceptance radius but that doesn't seem to be the way it works

[ocean wren]

Its the centroid of the capsule isn't it? might be worth checking the code though

[fluid cliff]

Hi guys. I might have some misunderstanding of behavior tree logic. Can someone please explain this?
This screenshot is from the original docs https://docs.unrealengine.com/4.27/en-US/InteractiveExperiences/ArtificialIntelligence/BehaviorTrees/BehaviorTreesOverview/
What I do not uderstand here is the FindRandomPatrol node which is done in a sequence. If behavior tree is always executed from root to all(some) of its descendants wouldnt BTT_FindRandomPatrol be executed every time? Wouldn't that end up in picking random point every time while moving or waiting ?

[flint plover]

I'm having a problem with Nav Links, for simple testing I have 2 rooms with Nav Volumes and both are connected by a simple Smart Nav Link. I have the AI Character in one room and a target in the other. So I do a simple AI Move to target. Anyway I would assume the AI would walk up to the Nav Link sense that is what connects the to Nav Volumes, However the AI decided to walk to the wall, and what is even more bizarre is the AI Move To results say it was successful despite the AI being no where near the target. I have no idea what is going on here.

[uneven cloud]

The BT is evaluated left to right. Selectors run tasks until the first one returns true and sequences run tasks until the first one returns false or all have been executed.

In your example, it tries to chase the player. If all of the tasks complete, it will return to the root and restart. If the blackboard is false or any of the tasks fail, then it will patrol.

[uneven cloud]

Is there an arrow connecting where the smart link should be when you show navigation? Is the wall blocking navigation? What does the visual logger show the path to be?

[flint plover]

do you mean the green arrow? yes... and yes wall are blocking, I am also using 2 Navigation Volumes with the nav link connecting the 2

[silent hamlet]

There is, if you check in the transition section you'll find it. It used the gameplay tag system

*uses

[rancid wing]

so for making the enemy see the player, do I put a perception in the charcter bp?

or is there a more effiecient way of doing so? cuz I have the bt and bb and all that

pls @ me

[flint plover]

So made a fresh project just to test if the issue with the Nav Link is present, and turns out it the problem is present in a fresh project. (the Cube AI is ment to go to the Nav Link to get to the other area that has the Empty Actor (White Sphere)

[tiny comet]

Thank you ! I never would have found it

[uneven cloud]

Better to put the perception on the AI controller.

[rancid wing]

oh huh

bro chatgpt told me the exact opposite lol

ill try that out

[uneven cloud]

Do you still see an arrow when the nav link isn't selected? What does the visual logger say about the path it's attempting? Is it a partial path?

[flint plover]

The arrows only show when I select them

[flint plover]

this is what the logger showed me

[uneven cloud]

Then the link doesn't actually exist. Likely because it's in 2 different nav volumes, but I would verify that.

[flint plover]

I see, yeah I open up one of my past game projects and the Nav Links in that game fully show the links without me selecting them

[rancid wing]

and does the ai perception need navmesh?

its not detecting my player

[flint plover]

Though it can't be that its 2 different nav volumes cause the other game that works is using 2 nav volumes

[rancid wing]

I have everything setup as well

like the stimulisource and what not

nvrmind stupid fix

be sure to check nuetrals and freindlies ig lol

[rancid wing]

so, how do I get a players location with ai perception and make the ai rotate to there?

[terse panther]

when the MoveTo or wait node are executing if kinda stops all other tasks in the BT, so after picking a random point it will execute the MoveTo, and then it will wait unitll the Pawn is moved to that point, and then the Wait node will execute and after that it will rerun the BTT_FindRandomPoint

[sudden citrus]

Has anyone had experience using Kythera AI for Unreal and is willing to share their impressions?

[celest python]

If you dont make it obvious that you are very rich, they dont answer your emails

Customer service is beyond terrible

AI tools are quite great, but you dont work with UE workflow. There are weird way of doing things and making you get confused

I cant share the specifics because of NDA

If I would ever end up having that amount of money to waste to Kythera I would just hire a team to make something New instead, at least I wont bother with some weird web api based behavior tree and unanswered emails when I need support

[sudden citrus]

Oof. Thanks, that's not what I expected to hear

[misty wharf]

Hrms, I wonder what would be a good way to communicate results from smart objects back to the actor interacting with it...

I guess if it's a BT based SO it can just have some BT node in it which calls some functions on the actor 🤔

Seems that if the BT aborts it doesn't count as a failure for the SO either

[celest python]

I really enjoyed their BT editor, but web api and customer service was a huge deal breaker

not to mention weird ways of communicating with web api from UE editor

[misty wharf]

Sounds pretty funky to have a web api thing like that... but on the other hand it seems it's probably a lot faster to iterate some stuff like that

[celest python]

its to ship to different engines at the same time without custom editor tooling for each

[misty wharf]

Yeah

[celest python]

technically I didnt sign an NDA, my client did, but I guess I could risk him if I explain more 😄

[misty wharf]

It's a bit weird they'd be so secretive

It's almost as if they don't want more customers

:P

[celest python]

from what I heard from my client they make you sign NDA for each time you talk to them

( )

[misty wharf]

Somehow I'm getting the feeling it might not even be legally binding

Ugh, I really don't like blackboards, they feel like it's just a bag of random variables

So I think to communicate back to my actor from my BT-based SO.. I'll just have some flag on my blackboard which is set to true/false based on whether the SO succeeded :P

[celest python]

This summer I'm planning to rework BT editor and blackboard system if I can find time, and will use plain data assets to define blackboards.. I have no idea why we need UBlackboardKey while reflection system exists

[misty wharf]

If you can just implement the ability to reorder BB keys that'd already be good

lol

I already have a variable called just "Bool" which is used for random purposes lol

Oh, at least there's entry category... not sure if that was there in 4, at least I never paid attention to it back then. Makes it a tiny bit better organization-wise I guess :P

[sudden citrus]

as a PR or are you at Epic?

[celest python]

just a custom source build for myself

[sudden citrus]

the thing that's bothering me the most is dynamic subbranches only being allowed to use the parent's BB type

essentially the opposite of a callstack

[celest python]

My idea is removing dependency to AIController and adding more modular composites to allow scripting in BT editor directly, and a better message bus system to get callbacks from BP to not write a C++ task each time I need to know if a blackboard value* is changed or not

[sudden citrus]

each time I need to know if a delegate is changed or not
what do you mean?

[celest python]

blackboard observers

[misty wharf]

you should introduce a new type called Chalk

which is used to write to the blackboard

lol

[ocean wren]

I think the issue was that the BB predated the property system from blueprint? or that there was some perf concern? I'd have thought that handling property change events would be the way to go and not even have a BB at all, but instead have the blueprint's variables accessible to the BT editor.

But on the other hand, we really need shared blackboards across agents.. which is why they were invented really.. So maybe have a system that allows you to observe another actors property changes too

[celest python]

I would guess BB predated property system, or simply just it was a bad idea. There wouldnt be a perf difference at all between reflection access and current implementation

I think ideal BB system should allow dynamic property adding and removing too, like an array of arbitary properties

[ocean wren]

Yeah, it doesn't seem particularly useful to have a different UI for BB considering what the BP editor has in terms of tools etc

Yeah, exactly

I mean basically, its the blueprint variables we need 🙂

And some way of synchronizing those across agents (i.e. sharing of BP variables)

I always just wanted to have a TArray of enemies for instance.. but BB doesn't have arrays.. seems silly

[celest python]

since you like slate, there is a "property access" module in the engine, i guess you can find it useful for your future experiments

its basically a generic data binding system that used on ST and anim graph

[ocean wren]

Oh? niiice!

[misty wharf]

never used shared BB stuff even though it has that option

not entirely sure where that would be useful 🤔

[ocean wren]

Slate and the Editor framework in general are pretty well designed.. aside from a few issues that if I were working on AAA stuff I'd have no problem fixing

The shared BB stuff isn't currently really useful, because you can't specify who shares it

imagine a squad having squad members with squad-member-only BB values

[misty wharf]

Right

[ocean wren]

but also having "all squads" for command of the overall battle

its a way to have different channels of data really.. like a pub/sub kind of thing, but for data

[celest python]

also see this, it should have some cool examples: https://github.com/DoubleDeez/MDFastBinding
this plugin is example of how current engine implementations can be improved imo, UMG bindings are terrible in terms of performance and this one using reflection access to add fast data binding

[ocean wren]

Hahaha.. so many githubs I've starred for later 🙂

later never comes

Actually, thinking about it, you could use the property system to connect agents via a component, just have the component respond to property change events and propagate them correctly

[celest python]

Can you elaborate more

Like having a component as manager simply?

[ocean wren]

yeah.. essentially a publisher/subscriber thing, but for BP's to coordinate their variables

think about it, with reflection, all you would need is public access to whatever you wanted to share

and events for when something changes to cause propagation between agents

[haughty coral]

What I do is that AI Controllers have pointer to the shared blackboard (custom UObject with BP properties). The pointer is copied to ST and ST tasks then refer to squad shared entries over it. For example: Squad Blackboard -> Target

[ocean wren]

So say you have a master actor that has all the shared data.. with events for requesting changes.. and then "subscribers" that can access that data as if it were local by subbing to the master actor

Yeah, I've used a shared squad actor as data holder before

but it feels like it could be a nicely done UI thing to have like publisher/subscriber data access as components rather than requiring another actor

[haughty coral]

The blackboard does not need to be a component. Can be just UObject defined without any actor actually owning it.

[ocean wren]

I'm thinking in terms of usability and UX here

from the Blueprint and BT, I'd like to just access data as though it were locally defined, no matter where it comes from

[celest python]

I'm thinking about letting BT editor take bindings from a data-asset which is a BB, and there is a global AI manager responsible of syncing global BB values

[ocean wren]

Make some wireframes so you can figure out the UI beforehand, rather than just throwing it together 🙂

I'm terrible for the "throw it together" thing 🙂

I guess I like to understand problems by facing them and then figuring the solution

bbiab, got a student meeting

[haughty coral]

Would be also good to not expect anything being global. The BB syncing in UE being completely global is also one of the reasons why it is not really useful. For example, BB can be shared with a squad, on the higher level there can be a BB shared with a faction and one higher level there can be one shared between factions.

[celest python]

You're right, maybe having a group system is better, and letting AIs participate into multiple groups

so user can create a group for "global" access

and local groups

[haughty coral]

yes, that's what I do

basically just a generic concept of groups which share knowledge / BB

[celest python]

nice

kythera was also same 😄

[haughty coral]

I tend to use wrapping gameplay behaviors which spawn a child gameplay behavior and track its lifetime. In your case it could be a gameplay behavior which waits for the child behavior to end and then does some custom notify.

I use it for Time Limit gameplay behavior which basically just limits the time child behavior can run. Another use case is sequence behavior which runs multiple behaviors in a sequence.

[misty wharf]

Interesting 🤔

For me I have AIs which have sets of tasks and items they carry around, and using a SO may add additional items and potentially affect task state. Currently they're just handled directly in the main BT's for it, but I think I can probably just adjust it for SO's

[haughty coral]

I see. Depends if you are 100% it should always happen when the BT ends. Otherwise it might need to be handled within BT. In case of interaction stuff I can see some effects requiring being applied mid-iteraction (for example when character touches something with hand) and then the iteraction just finishes with some cosmetical blend out animation.

[misty wharf]

Yeah.. I currently handle a lot of it directly in the BT, but I'll probably make some AI Tasks instead for it which can be called from the BT tasks. Might make it easier to handle errors like the item being unavailable (eg. because someone else took it or such), since it's a bit of a hassle to categorize different types of error scenarios in BT logic

[misty wharf]

Hmm looks like there might be yet another feature that I think is not very well supported by the builtin SO's... I might want to adjust the activity tags at runtime 🤔

However it would appear the activity tags are only defined in the data assets, so it seems if I was to adjust them at runtime (surprisingly there's a function for it), it seems that would change all of them that share the same definition...

Maybe I'll just build a separate filtering phase on top, where the activity tags on the SO define all possible choices for the object, but the actual currently available tags are stored separately. At least it'll allow it to reduce the number of objects the query will return

[gentle sail]

HI , for start I am French , sorry for the approximate language , I work on a 2d project for 2/3 days and I try to create an ai wo should be able to move and I have add a nav mesh bounds volume but when I start the game , the ai dont come to me, I have try a lot of times. Help me please. note that the floor is made of paper sprite, (I don't know if that changes anything)

[haughty coral]

I hit the same issue too :/ For now it was not hard requirement on my side, so I skipped it. If you are willing to update, you can check if changes in 5.2 help you with that. Slots have RuntimeTags property and also world conditions driven Preconditions for selection. Might be worth investigation.

[misty wharf]

Ohh interesting

I don't really want to maintain a custom engine build, but good to know what's coming. I'm kinda building mine as an abstraction so it'll hopefully be easy to adjust in the future if they add better support for stuff (or if I need to replace it altogether)

[haughty coral]

There are also funcs like AddTagToInstance in subsystem but I don't remember how these work. https://github.com/EpicGames/UnrealEngine/commit/c0176005fa6678513a42414dd2c9823ad1920d50

Check the FInstanceTagsFilter test in this commit. It seems it creates a filter for instance tags. This was commited year ago, so probably even in 5.1.

[misty wharf]

Not sure what that relates to, seems to have something to do with Gameplay Abilities 🤔

I was looking at the codepath that uses the activity tags you can define when searching for smart objects, but I can't find anything in there that would allow defining them at runtime in 5.1, so I guess this doesn't affect that

[haughty coral]

I think they intend activity to describe only the action type, basically categorizing it. Like SOActivityTag.Door.

And instance tag would be then containing its dynamic state like SOInstanceTag.Locked

in SO Def you configure the ObjectTagFilter that this object can be used only if it has no tag SOInstanceTag.Locked

and when locking / unlocking the door, you use subsystem funcs to add remove SOInstanceTag.Locked tag

but still it is kind of.. half dynamic approach because you don't construct the tag requirement query per search but define one per definition

Smart objects instances are internally represented by FSmartObjectRuntime. The AddTagToInstance gets that internal representation and adds the tag to the instance. So there is no GAS involved.

void USmartObjectSubsystem::AddTagToInstance(FSmartObjectRuntime& SmartObjectRuntime, const FGameplayTag& Tag)
{
    if (!SmartObjectRuntime.Tags.HasTag(Tag))
    {
        SmartObjectRuntime.Tags.AddTagFast(Tag);
        SmartObjectRuntime.OnTagChangedDelegate.ExecuteIfBound(Tag, 1);
        UpdateRuntimeInstanceStatus(SmartObjectRuntime);
    }
}

USmartObjectComponent is going one step further. If you use GAS, it keeps your SO runtime & ability system component in sync but you can still add or remove own tags manually via the subsystem. But it is not necessary to use GAS with the component.

Would be definitely nice if there was a way to define also InstanceTagRequirements in the FSmartObjectRequestFilter. However, FSmartObjectRequestFilter has

TFunction<bool(FSmartObjectHandle)> Predicate;

so you can configure there your own filtering function which checks SO Runtime instance tags. That should do what you need.

[misty wharf]

Ah, yeah need to look into that some more then 🤔

One wishlist item: I hope they improve the BT debugging facility

When it switches to the SO behavior tree it kinda breaks the ability to easily step through your BT to debug it which is somewhat annoying

This is already making it annoying to try and figure out why my BT's aren't currently working properly because I can no longer step through them properly... ugh

Oh.

So the smart object behavior tree... when it deactivates... it doesn't actually seem to resume from where the previous tree left off. Is this for real?

:P

I knew there had to be some gigantic gotcha in this because it was going to be too convenient otherwise

Or maybe it's a problem because I'm using a subtree and it doesn't handle that correctly...

Nope! Still don't work if no subtree is used. Who even designed this and thought this was a good idea...

[lyric flint]

I want to attach the stimuli source to a component in my BluePrint with many components. Is that possible? or do I have to make an entire new blueprint, register that as the stimuli source and attach it?

[crimson oxide]

Hey there; any tips on how to make different AI agents to communicate with each other, for example if one enemy detects the player then that knowledge has t obe communicated across AI agents. should I include this in a service ?

[misty wharf]

There's a variety of ways of doing it. You could for example use a shared BB key, or use some mechanism such as finding all other characters within a radius and calling some function on them

[crimson oxide]

sorry how do I set a shared bb key ?

[misty wharf]

If you click on the key in the BB editor, it should have a checkbox which allows sharing it

@haughty coral good I saw you typing, I was gonna ask if you had any idea regarding my above gripes with the SO behavior tree stuff lol

[haughty coral]

Oh yes yes, that's the main reason for my recent rants about BT & BrainComponent in this channel and also why I was investigating running multiple BT Comps 😅 I am afraid I don't have satisfying answer. There can be only single BT Comp meaning the behavior needs to nuke the previous one and then pretend the old one continues by starting it from the root again.. what of course does not lead to desired behavior in multiple cases.
Previously I did BT saving by subclassing BT Comp and doing a big hack fest to trick it into resuming from a specific node but that was just a horrible thing.
In this project I decided to exchange root AI logic system to State Tree. So behavior's BT Comp does not need to fight for Brain Component. Brain Component is always used by behavior BT trees and the main State Tree based logic is handled outside of brain component stuff.

[misty wharf]

Interesting, so you can run BT and ST side by side like that?

I didn't look at it much yet, but was considering using ST instead of BT for the main AI logic since it might actually fit the paradigm better anyway

[haughty coral]

yes, ST is just a structure so you can run it within anything and even multiple STs in parallel.

[misty wharf]

And/or wrap logic in AI Tasks that I can then use directly in Gameplay Behaviors in blueprints that are safe to call in BT's

Although going to need to make some kind of mechanism to communicate the results back from the gameplay behavior since looks like it doesn't have anything for it builtin, but probably just gonna add some delegate to my character class like OnFinishedGameplayBehavior or something which can then be used to get a result of some variety

[haughty coral]

I guess it depends what the results are. I would hesitate to do anything with UGameplayBehavior objects besides just triggering it and waiting for its end. But If you really need the result I guess something like GAS gameplay event / lyra message / delegate is the way.

[misty wharf]

Hmm yeah I guess I could probably verify the result of the behavior after executing it, instead of using a result value to determine it

One behavior for example would be that the NPC picks up an item from the shelf, which either succeeds or fails, and the NPC's task object should be marked as completed or failed depending on whether that behavior succeeded

[haughty coral]

It really hurts me when I keep thinking how much cooler BTs could be with only a few smaller touches like detaching them from BrainComponent or turning BrainComponent stuff into an interface instead, so that we could choose to instantiate BT implementing BrainComponent interface or a completely separate standalone BT instance.

[turbid reef]

CAN SOMEONE HELP ME WITH THE AI OF MY RACING CAR, I CAN PAY. PLEASE DM

[misty wharf]

This could certainly also be verified simply by looking if an additional item of the correct type was added into the NPC's inventory after the gameplay behavior has finished

Hah yeah, another issue I had with BTs is that they really really hate savegames... probably the same reason why it can't resume from where it left off with the Smart Object

If it only was able to serialize its current execution state, but I guess it never occurred to them when they designed it originally. That would make it possible to both save it, and resume it if it has been paused for some reason such as for a smart object

[harsh storm]

Just don't let people save their game 🧠

[misty wharf]

That's... actually partially how I'm solving it in this project

I only allow saving when it's not going to be a problem lol

[haughty coral]

Gameplay behavior can either finish successfully or abort. You can bind to a delegate of specific gameplay behavior instance to listen for the latent end and get the end result (success/fail) from there. That's what I usually do.

There is however one weird thing to keep in mind. The return value of behavior's Trigger function does not mean if the behavior succeeded or failed immediately. It means whether you can bind to behavior end delegate and wait for it being executed at some point. This has big downside.
You can not fail the behavior on startup with just returning false as false can also mean the behavior finished successfully immediately, so there is no delegate to bind to. So if you want to fail the behavior, you need to set timer on the next tick and abort it from there 🙃

[misty wharf]

Well that's a nice design lol

[haughty coral]

Yes, that's exactly the reason why I implemented BT state resuming on the previous project. To my surprise, it was possible only by inheriting BT Comp. But I really had to tinker with it for days to set all the variables correctly to really trick it to start from a specific point. Most of the time it was just trial & error where error means crash haha...

I hope ST is better in this, did not check

but as it is just a structure representing its state, maybe it is enough just serialize the structure

[misty wharf]

Yeah I figured the data to load/save is probably in there somewhere but never saw a very obvious way of doing it when I looked at it, so trial and error sounds about right :P

Thanks, your insights have been very useful for this

[bronze wagon]

Kaos, I tried setting the ObstacleOnly last night and it didn't work as I intended. What I'm trying to do is if an AI is idle, I can switch them to an obstacle so they are routed around. Is that possible?

[gentle sail]

HI , for start I am French , sorry for the approximate language , I work on a 2d project for 2/3 days and I try to create an ai wo should be able to move and I have add a nav mesh bounds volume but when I start the game , the ai dont come to me, I have try a lot of times. Help me please. note that the floor is made of paper sprite, (I don't know if that changes anything)

[bronze wagon]

What's the movement result?

[gentle sail]

My ai don't move

[bronze wagon]

The value of the MoveTo result

[gentle sail]

My artificial intelligence is supposed to follow the character

[bronze wagon]

There is an output of the AI MoveTo in your screenshot that says "Movement Result". You need to see what that result is

[gentle sail]

I don't think it's important, I have don't see this in turorials

[bronze wagon]

ok then

[bronze wagon]

So I have checked Visual Logger and CrowdFollowing is being set to ObstacleOnly on begin play for 3 characters I have lined up to block a doorway. The other characters are all Enabled. They are still trying to path through the ObstacleOnly characters. Using the debug view on their pathfinding they are not registering as blockers.

Ah-ha, I enabled CanEverEffectNavigation on their capsules and it seems to be working

[final loom]

So is Get Blackboard used within custom tasks?

[lyric flint]

It is important and you should see what the result is

[daring fox]

Is posible to integrate chatGPT in an Unreal game?

[misty wharf]

If they provide an API you can use, yes

[crimson oxide]

Hey there, is anyone aware of the most performant way to get the actors of a class within a radius (?

[bronze wagon]

There is an integration plug-in on the marketplace

[daring fox]

Thanks

VaRest plugin?

[bronze wagon]

https://www.unrealengine.com/marketplace/en-US/product/chatbot-ai-unleashed

I bought it the other day picking up some other assets. Haven’t implemented it yet

[uneven cloud]

What is it that you are trying to do? Stimuli source component is not one that you attach.

[uneven cloud]

Get actors of class is not actually that bad in a packaged game, unless you have thousands of them. A more performant way is to store them in a different data structure than the TMap the engine already does.

[ocean wren]

https://youtu.be/oiyFTYQ07Vs <--- mine

Its just a HTTP rest api, pretty simple to integrate (chatgpt needs a set of message strings to initialize with)

[daring fox]

Thanks!!!

[daring fox]

This is awesome, congrats

[daring fox]

Let me know if it works when you implement it please

[bronze wagon]

I started to but I signed up for the chatgpt API 6 months ago and didn’t know my credits would expire. They expired Dec 31st so I need to pay for more. Will probably be a few weeks before I get back to it. Also the API doesn’t appear to support the same chat style interface, it is focused on completion.

[lyric flint]

@uneven cloud I figured it out. I put a stimuli in a far away part of the scene and when the ball spawns it'll attach itself to the ball. it was a pretty simple fix that I figured out. I appreciate you trying to help out.

[crimson oxide]

Guys, im assigning a Blackboard key in a service, but then I realize that the targetKey resets. How do I assign a Key a value from a service correctly (?

[daring fox]

Ok, understood🤘🏼

[quaint radish]

does a pawn need to have the parent class "eqs testing pawn" to use eqs?

[harsh storm]

No

[quaint radish]

Any way to query a volume of space instead of just a horizontal plane around the querier?

this for example will just generate points on a plane

I am dealing with flying ai and would like any ideas for checking z or verticality

[uneven cloud]

You can make your own generator.

[final loom]

What's the most memory and performance efficeint way of passing the name value? A constant, or just make name at the time that gets set?

[misty wharf]

It's unlikely to have any effect on performance whatsoever regardless of how you do it

FName values are basically converted to numbers if I remember correctly, so there's virtually no impact from it on memory use

[final loom]

That's helpful.

Do you know how I add properties to custom tasks/services?

I hope me using custom tasks won't bog it down too much...

Can I make services run only once somehow?

If I set it to 0, will it run every frame?

[patent hornet]

watch the WTF is? BT Service youtube video

https://www.youtube.com/watch?v=49Xub-taTig&t=8s

[final loom]

Huh. I may have been doing it right with using on activation.

Ohhhh. I needed to make the variables public. Gotcha!

Glad to hear I can do that though

[neat gate]

I am throwing a rock and invoking Report Noise Event. My AI is listening to the OnTargetPerceptionUpdated event, and it never fires. The AI has the hearing stimuli registered, and its range is 3k.

Why doesn't the AI hear it?

[lyric flint]

I'm having an issue with the MoveDirectlyToward and MoveTo task in the blackboard. Whenever the pawn senses the target actor it's suppose to track, it'll then track it extremely slowly and not use the acceleration and full speed in the movement component I have for it. Is there any way I can override those tasks or do I have to make my own MoveDirectlyToward and MoveTo tasks?

If I do have to make my own tasks like that then does anyone have a video guide I can use? I'm still pretty new to blackboard and want to be able to do a lot with it.

[lucid marten]

In UE4 I am having a issue where the AI is following the player but not attacking the player after spawn. When I put AI bp in map without spawning, the AI works perfectly well but after spawning AI is not attacking player. There are two halves in behavior tree where in first half AI is instructed to follow around the players and in second half, AI is instructed to attack. First half is working but second half is not working.

[tawdry crag]

How do I change the agent index of a navmesh agent?

[stoic pier]

With EQS how can you check if a node is already occupied by another AI? Trying to avoid 2 going to the same place

I have a visibility test which works for walls but placing another AI actor in the way doesn't effect the filtering / scoring.

[misty wharf]

If the AI somehow "reserves" the node it's going to, you could use a custom EQS test to filter out nodes which have been reserved already

[stoic pier]

yeah, trying to figure out how to filter those, probably with a task

[tidal seal]

Hi guys, I'm struggling to make RunDynamicBehavior node works, after checking quickly the source code, it seems that my blackboard is incompatible, I guess between the MainTree blackboard and subtree blackboard, am I forced to duplicate all the entry in the main tree blackboard in order for them to be compatible ?

[little basin]

Hi, does anyone knows why a custom decorator doesn’t tick even when I have the bNotifyTick set to true? I find that confusing

[uneven cloud]

With the noise events, what are you sending for the sender? Since you are listening for On target perception updated, it's likely you are listening for the player and sending the rock.

[uneven cloud]

You do not need to override or create your own move to. They do nothing about the speed. The issue is with the movement component.

[uneven cloud]

What do you mean? What are you trying to do?

[uneven cloud]

Since the BT sounds like it's running, I would check the vis log for why the attacking branch is not running.

[uneven cloud]

If visibility is not working, then you might want to verify the actor is blocking visibility in the collision settings.

Another easy way to do this is to create a context that gets the other AI agents locations. Then do a distance check using that context.

[tawdry crag]

I wanted to have navmesh links selective allow some agents along them but not others. I hacked it in in the end, as AFAICT the only way to do that in UE's implementation is to use multiple agents defined in the project settings, but then you get multiple navmeshes and the assignment of agent Indeces is hot garbage, so hacks it is

[uneven cloud]

Nav areas and filters would be a much cleaner and easier way to do that.

[tawdry crag]

Yeah, but the nav areas only filter by agent types, right?

Oh, the query filter can filter by nav area?

OK yeah that would probably have worked

[uneven cloud]

Nope. You make a nav area jump and assign that to your link. Then create a filter to exclude the nav area jump and assign that to the default filter for the AIs that you don't want to use them.

[tawdry crag]

So that's not a query filter?

[uneven cloud]

It might be called a query filter, but I don't remember the exact name.

[tawdry crag]

NavigationQueryFilter is the only kind of filter I can see

I was just trying to understand why you said "nope"

[uneven cloud]

That must be it.

[tawdry crag]

I think we are saying the same thing

[uneven cloud]

I said nope to your first line. About nav areas filtering.

[neat gate]

In the unreal tutorial video they didn’t set anything as the sender.

As I’m doing this in code now,I figured I should send the owner of the rock. Correct?

[uneven cloud]

Or the instigator. It depends on what you used for spawning.

[tawdry crag]

The areas themselves only care about agent IDs, but I see what you mean

[neat gate]

Ok give it a try when I get home

[night relic]

Should I get a significant performance gain by using navigation invokers rather than a navmesh in an average 2x2 KM level?

[uneven cloud]

No. Invokers are far more expensive.

[night relic]

I thought that was the point of using them

But the documentation is ambiguous as fuck, so you are maybe right

Navigation Invokers are ideal for big Levels where building the Navigation Mesh in the editor is impractical.

I thought it was a matter of performance

[celest python]

how can I learn what is blocking AI's movement?

[bronze wagon]

I’m currently digging in the C++ to get better handling of the blocked status. So far not much progress.

[celest python]

turns out it was a GAS attribute error, MovementSpeed attribute was zero and CMC was taking it as input on GetMaxSpeed

among tons of possibilities there wasnt any easy way to detect that

[bronze wagon]

That’s most of the path finding problems I run into it seems.

[crimson oxide]

Im having an issue when assigning a blackboard value from a service, thing is that it seems like it resets to noone when the actor is marked as alive, whenever the actor is killed the blackboard doesnt reset for some reason

Almost looks like some kind of bug because leads in my behavior tree are executing even tho they are disconnected

[dusty merlin]

Hi everyone, I would like your opinion about how you would use a Reinforcement Learning framework (something like RLlib) to integrate into your game or what stopped you from doing so so far. I know there are some official tools for other game engines, but I didn't see much for UE. I'm building a framework for putting RL into games with minimal effort for the user, and I would like some feedback from anyone who spent some effort or who does not know anything about AI or RL but still wants to implement some features. It will be beneficial to decide which features to develop and how to make it as easy to integrate as possible.
Thank you!

[night relic]

Is it a C++ project?

[empty dragon]

Anyway in the new state tree system do print string debug out the current state ? Can't seem to find the getcurrentstate node? Im having trouble with the Completed / Succeeded state transitions

[stoic pier]

cool, thanks. I'll check the collision. Ultimately I think a manager actor that runs EQS and passes the locations from an array might be the best bet. Still have yet to test this tho.

[glad stone]

Hello everyone. I'm trying to make AI controlled car from vehicle UE5 template. Car has vehicle movement component. But in Behaviour tree task Goto not affected to the car. Any suggestions why it not working? Thanks everyone!

[sudden citrus]

Are there any plans at Epic to implement automatic navlink generation? Any advanced traversal support at all?

[ocean wren]

Doing my own.. the main thing for me is the user experience. Having to step outside your knowledge of the toolset already established is a no-no. Look at Unity ML agents as a good example of how not to do it.

Personally, I'm more interested in scaling up and productionising the whole RL process.

[cinder condor]

Question: I've seen the videos pinned, and they sound pretty awesome. Do we have an example project or diagrams of how to achieve a good separation of concerns within UE's AI functionalities?

[sudden citrus]

@crystal hatch I'd like to generalise most of MassActorSpawningSubsystem down to a base class that lives in the AI Module, with two child classes: one for Mass, another for plain actors. Doesn't look like a lot of work. Was building my own spawning subsystem and realised I would just copy your code bit by bit... think it's a good PR?

[worldly shard]

In 5.1 StatTree, seems SoftClassReference as parameters broken, they are always invalid in BegingState, anyone any experience with this? Only happens when you test in standalone

[celest python]

i wouldnt be surprised

its quite buggy about properties

[worldly shard]

I see! I'll work around it for now 🙂 thnx

[uneven cloud]

FYI MassActorSpawningSubsytem can be used for plain actors. I'm using it currently.

[sudden citrus]

yep I've just tried - I just don't like potentially interfering with the mass actor spawn queue/budget/etc, and it doesn't need to depend on mass

the core functionality is just really clean

[uneven cloud]

I just didn't want to write yet another pooling system.

[sudden citrus]

same

I'm gonna do a smaller PR first thought to add a delegate before the FinishSpawning call

it even says TODO add code here 😄

[north oriole]

How does navmesh work in relation to world partition?is the navmesh simply recomputed each time a cell is loaded (for that cell)?

[sudden citrus]

With static generation, it is streamed in chunks along with the cells. There is a separate navmesh building process for WP which saves the data out into hidden actors per chunk. Google "world partition commandlets"

Dynamic generation for WP is still experimental I think

Also your recast navmesh actor needs to be marked as "is world partitioned" for the above to work

It's all a bit jank atm cause the editor will also do its own transient generation if you move anything, by default

[north oriole]

Ok so if my game uses dynamic navmesh it would be better to avoid WP?

[sudden citrus]

Possibly for now, yes

It might come in with 5.2 but I've not been following, we just set ours to static and are waiting for news

[north oriole]

I'm assuming in your game you don't have stuff like locked doors that you can unlock and so on, because for that you would need dynamic mesh necessarily i think

[sudden citrus]

You can do static with modifiers for that

I think smart navlinks also work with static nav

[slow bobcat]

I can confirm smart links work with Static nav (both pure static and dynamic with modifiers only)

[sudden citrus]

Doors are arguably better solved with modifiers, then you don't have the problem of "how do I place navlinks along a really wide door"

but for normal room doors you can do a navlink, it also gives nice functionality like access checking - they can be locked but checked during pathfinding and unlocked during traversal

don't @ me about performance 😛

[slow bobcat]

BTW, would be nice if there was some clean-up with smart links. Setting them is not great:

• Set a proxy nav link
• move the Left/Right
• Copy points to smart link with the button in the details panel (callInEditor function)
• Set smart link to Relevant (whether is enabled or not)
• Remove the simple-link points (delete array). Otherwise your smart link will never use the smart link features.

You want to edit the left/right locations again? well... no button to copy-back from smart to simple and only the simple link has the Left/Right gizmos on the level to easy-edit the locations.

Not great.

[crystal hatch]

Might be interesting. If it’s not too much work then post the PR. Worst case scenario we won’t pull it in and you end up with a generic spawning approach that you could make into a plug-in and sell on the marketplace;)

[sudden citrus]

Yes and no, if I turn the mass version into a subclass and you don't take it, I'm stuck merging engine upgrades manually 💀 I'll get to work then

[dusty merlin]

What was the problem with Unity ML agents?

[empty dragon]

In the new state tree system, how can I use a notify event in an animation montage to change the state? Obviously I could have a bool that tracks the change and then use an evaluator to wait for the bool change. But is there a way to directly link a notify event to effect a state change?

[dusty merlin]

Should it be like a plugin or a more specific integrated tool?

[ocean wren]

They had an external python program you had to run to actually do the training, which wouldn't be so bad if it were a cloud service, but sucked to use otherwise and was fiddly to get working

[ocean wren]

Personally, I'm going for a mixture of plugin and components.. So a component that "owns" a data asset that describes the actions, state etc. Essentially a connection between properties on the thing being trained.. but mine will actually go out to a cloud service for the training

[celest turtle]

anyone know how to add more data into FEnvQueryResult ? For example I would like to add an normal of a trace hit. From what I'm understanding the result is only storing actors and vectors as locations OR directions. I would like to have an vector as location with additional vector with directions.

[north oriole]

Is static with modifiers supported and dynamic no?i encountered a few bugs with static with modifiers

[bronze wagon]

Is there a performance benefit for EQS vs GetAllActorsByClass?

[celest python]

if you run C++ version of GetAllActorsByClass, it should be performant enough - i doubt the implementation inside of the test does something different than just using TActorIterator

but compared to BP function yeah

c++ implementation is very smart but BP one just yoloing an expensive lookup and some branches to filter to given class and copies array to output param in VM

[sudden citrus]

EQS is not a replacement for GetAllActorsByClass, in fact you can use the latter as a generator in EQS as Eren said. EQS is a way to filter those actors, score them, pick a winner. Or to do the same with generated points. In a data-driven way.

If you want to be quicker about finding the actors in the first place, use an octree or a collision shape.

Having said that, if you have very few actors of a given type, then GetAllActorsByClass in C++ will be quick enough because it's bucketed by actor type

[bronze wagon]

Cool I’m using blueprints to call it now but I have some things I’m doing with C++ for pathfinding so I’ll add some code based lookup functions

Thanks! I’m working in smaller levels so it isn’t a ton of instances I’d be returning. I just loop through them to find the closest one or pull Random on some.

[bronze wagon]

I have a very frustrating problem with AI Move To. I have characters sitting in chairs. When I want them to stand up and move, I have to add .2-.5 second delays on either side of changing the collision type or they will be stuck in Blocked status indefinitely. With the delays long enough (nearing 1 second), they will move fine. About .7 seconds and the first attempted Move To is blocked, but the next one works. I have verified all of the code running the standing/collision setting completes before the MoveTo is called. I have also tried adding the collision setting via C++ just to see if it would help, and it doesn't matter. This means every time I want a character to get up and walk, they have to stand for 1 full second afterwards before they can move. What's up with that?

[bronze wagon]

Well sometimes it’s easier than you make it out to be. I’ve been fighting that issue for a while now then just tried re-enabling collision before standing up and it works smooth with no delay.

[bronze wagon]

And that didn’t work for long. Been broken since about 10 mins after I typed that.

[misty wharf]

Wait so when the character sits you're toggling collision settings on the character or the chair?

If you're using dynamic navmesh this is likely to cause the navmesh to regenerate when you update those settings, which is likely to invalidate move requests near the affected parts of the navmesh

[bronze wagon]

I'm not using dynamic navmesh. Restarting the editor seems to have resolved it :/

[static crater]

Hi, looking for some global AI pathfinding advise. I got my own custom A* Pathfinding system. However i found its pretty slow when it comes with costs, like spline roads. Now i thought to create my own spline pathfinding. I think this is cheaper than A*. Anyone knows some terms to look for on google?

I think its somehow the same as car navigations. E.g. the pathfinder in GTA5 creates paths within a split second all around the map. What algorithms would be able to do so?

[buoyant geyser]

@lyric flint I'm HŌRU @HoruGame on twitter, we had a short exchange on twitter recently about recast and its performance. I actually have an idea that I think would benefit Unreal if you're interested.

[harsh storm]

Mikko hasn't returned since he first showed up GB 😭

[celest python]

Mikko wasn't active much since after the first day he joined the Slackers, he might not see your ping at all

[buoyant geyser]

hm I'll message him on twitter then, thanks :)

ok done

[uneven cloud]

You should be using Ignore Actor While Moving instead of disabling the collision.

[uneven cloud]

Fun fact, the BP one uses TActorIterator as well.

[celest python]

Yep, but in BP realm we lose the advantage of using template param to filter actors by class inside of the iterator with more faster way since it does some uobject hash tricks afaik

[empty dragon]

Is there anyway for the new StateTree tasks to listen to events?

or do they just have to watch bools that listen for events?

[celest python]

ForEachEvent or something

from context param

there is no observer pattern so far

[bronze wagon]

Hadn’t seen that before. Will take a look

I’m not disabling all collision, I’m changing the object type to one that the furniture and other characters don’t collide with. I have a group of people and they sit on a couch. The couch has a coffee table in front of it. I’m moving the mesh component out of the capsule to the couch sitting position. The capsule then also doesn’t collide with other character capsules when they need to move between the couch and the table.

[spring dune]

I have a door that can be opened/closed and a navmesh smart link that should be enabled/disabled depending on the door state. However, when I use the Set Smart Link Enabled function to set it to disabled when the door closes the AI still tries to go through the closed door. Anyone knows why?

[ruby shard]

why dynamic navigation in unreal is so messy ? it differs if i use component or static mesh actor

[lapis fractal]

Why is there no navigation mesh in the build of the game? The logic works, the problem is specifically in the absence of a navmesh

[lyric flint]

Is there way to have multiple pawns have the same AI. Kinda like Unities Entity and DOTS systems

[cinder condor]

What's the general opinion on State trees vs BT for AI? Any short list of pros and cons?

[celest python]

ST is a state machine with a selector on top of it, BT is a prioritization tool

BT = Systemic/sandbox behaviors
ST = yolo

ST is empty, BT is full of tech debt, both sucks

[cinder condor]

And do you know of any good locations where ST might be useful? Was thinking of a schedule (which is also shown in the docs) based on which the AI do different things.

But I'm trying to find a generic way to give an AI some specific tasks, and have those tasks somehow decoupled from the main BT / ST, and able to implement them in multiple iterations without touching the rest of the system

[celest python]

BT = I want an easy, readable, maintainable execution flow to prioritize behaviors, but I understand it might be limiting after a point and I cant jump between arbitary nodes based
ST = I want to be able to have top down decision makers and I also want to be able to jump from a node to a completely random node and I also care about cache efficiency a lot, I want to go yolo with tightly packed memory structs and slightly more painful workflow and I want to use something different than AIController and its components to bind data to my tasks/states etc

people here embed ST into a BTTask often

[lucid marten]

I am using AI character in my wave game without AI behavior tree. It has a problem where after player respawns after its death, AI no longer follows player. It just stays idle without any movement. How can I fix this issue? This is my blueprint of AI for patrolling around player.

[misty wharf]

You probably need to have it update the My Enemy variable when the player respawns

For example in the check where you test if it's still valid, when it isn't valid have it try to find a new target to assign to it

[pearl fern]

Seems like my navmesh is not present (or partially missing) in packaged build compared to PIE (unreal 5.1). Is this known? 😮

[lapis fractal]

Same problem

[prisma loom]

Im not sure i've understood BehaviorTree's and the Blackboard values correctly, I have a Tower that searches for enemies and if it finds an enemy it stores it's location in a blackboard value. Is this instance of the behaviortree and blackboard and it's keys/values not only created on that instance of the tower? Because when I spawn multiple towers they all seem to be controlled by 1 central BehaviorTree / Blackboard. Is everything shared or instanciated? Im confused

[lethal island]

Hi, I've set up a behaviour tree to create a dialogue system similar to older bethesda games (fallout 3/oblivion)

I've also set up a system for characters to be teleported upon interacting with doors, but want to use a similar system to go straight from dialogue options to another area

When I select the eye dropper to set the transition target it seems to think it's a valid option, but then won't actually select it

I've got near identical blueprints set up on a component to work for doors and it seems to work fine

Currently blackboard is set to an object 🤔

[lethal island]

It also finds it in the dropdown but won't set

[misty wharf]

@lethal island if I understood correctly - the object you're trying to reference to is in your level? If so, this cannot be done. Assets (such as behavior trees) cannot have references to things in your level, as the asset itself is not within the level. It's a bug that it even shows up in the selector, it shouldn't.

If you need to reference something in your level from your behavior tree, you need to have some other way of identifying it, and then having your BT task look it up from the level via some method

f.ex. you could use a method similar to how FlowGraph does it, where you can assign a gameplay tag as an identification, and have some way of finding the actor based on the gameplay tag

[lethal island]

Ahhh ok

[lethal island]

Is FlowGraph someone who makes tutorials?

Or a plugin?

[misty wharf]

A plugin

[lethal island]

Oh this sounds interesting

[misty wharf]

Yeah it's a handy plugin for doing a bunch of things

[lethal island]

So I watched a video from the dev about it...towards the end he starts talking about other plugins for dialogue but didn't seem fond of any of them

[misty wharf]

Yeah I've not used flowgraph for dialogue trees but there's some folks using it for that too

I mostly mentioned it because it uses gameplay tags to identify actors in the levels but I guess it could help with the dialogue aspect of your problem :)

[lethal island]

I mean, trying to use the AI behaviour tree seems like a janky shit show. Before trying that though I'd had several other tutorials just completely fail me 😦

I'm surprised there wasn't something for dialogue built in?

[celest python]

There is commondialogue or something

Luthage was mentioned it once

FlowGraph is king though

there is also this https://github.com/redxdev/Supertalk

[lethal island]

Ah flowgraph doesn't seem to play nice with this recent version of UE

[lethal island]

Ah, supertalk says it requires knowledge of C++ to implement. While I know a little programming I have no experience with C++

[celest python]

yeah but its some sort of a tree i guess, text based one

and a powerful one because it lets you talk with BPVM from its text scope

[celest python]

worksforme.png

finally there is also this https://github.com/NotYetGames/DlgSystem

literally just a tree

but not easy to work as flowgraph or supertalk

[lethal island]

doesnotworkforme.png

[neat gate]

So I want to visualize the enemies' sight perception in game. Is there an easy way to accurately do that? I figured I would render a sprite and multi its length by the sight distance

[lethal island]

What version are you running?

[celest python]

try running gameplay debugger

[celest python]

since 4.27, it works at each version

[lethal island]

But I can't open the editor?

[celest python]

generate project files

open .sln

or if you have Rider open it with uproject

run the project

see what error it giving

[celest python]

it was a response to hardcodednumber

[celest python]

this is your answer to the screenshotsh

[celest python]

also it doesnt complain about FlowGrap

you have a different module named dialoguesystem causing compiler errors

[lethal island]

Dialoguesystem is just the project I'm testing out dialoguesystems in

it ran fine before I put the flow plugin in

and in the first screenshot it specifically says flow

[celest python]

its not an error

[celest python]

its letting you know you didnt compile the plugin

[lethal island]

Ok, well the instructions on github simply tell me to unpack it in the plugins folder 🤔

[neat gate]

I know about that. I want to render it during runtime. For the player

[celest python]

it'd be a mix between #cpp and #graphics question i guess, you can just build a mesh runtime with the sight's mesh data and assing a material to it or do other tricks that I dont know yet

people are more knowledged about this at #cpp / #graphics

[neat gate]

yea i didn't think about generating a mesh. But yea, I figured

thanks

[lethal island]

Tried following the instructions here and it still won't build them https://dev.epicgames.com/community/learning/tutorials/qz93/unreal-engine-building-plugins

[lucid marten]

how can I do that can you explain

I am using AI character in my wave game without AI behavior tree. It has a problem where after player respawns after its death, AI no longer follows player. It just stays idle without any movement. How can I fix this issue? This is my blueprint of AI for patrolling around player.

[lyric flint]

Wait, MeshNav is something happen in runtime right? And if i want to find a pathway at BeginPlay() in C++, i can't use it?
If my understanding is right, NavMesh is something have to be placed in the level in order to work, but if compare that to pure math struct it would be much more expensive.

For example, i have several hundreds of "room" here, and i need to pathfinding between all the red room that is connected by the green line, NavMesh wouldn't be a good idea for that, right?

[eternal halo]

If I have a direct ref to a smart object component is there a way for me to find and claim a slot?

[ocean wren]

Navmesh generation can be both static and dynamic. Pathfinding is a separate process that uses the navigation mesh, in both cases you'd use something like A* or one of its variants.

[quiet shuttle]

Hey, when I spawn AI Vehicles into my level it automatically creates AI Controller for each.. it makes a lot of AI controllers if I have many agents... Can I somehow assign all of them to one AIController with C++ when I call SpawnActor<>()?

[empty dragon]

If you have a bunch of actors using the new StateTree, is there a way using blueprints to debug print out that state they are currently in? Cant seem to find any StateTree -> get Current state node that works

[harsh storm]

Controller can only possess one pawn at a time. You'd have to have your own system to do it the way you want.

[quiet shuttle]

ok, thank You

[rare hemlock]

Hey guys. I´m trying to use EQS to position my enemy characters around the player character, without running to the same position. For example I´ve got 4 enemys and they should position themselves in a circle around the player. So I would have to share the EQS data. Can anybody push me in the right direction on how to share the EQS-data?

[uneven cloud]

You don't share EQS data. You get the current location and destination via a context. Then do tests against those locations.

[cinder condor]

That sounds like a fair / good idea for some usecases! I'll check it out!

[rare hemlock]

Could you please go into more detail about that. Where would you check location/destination? In a Service? Are there bp nodes to get this data from context? At the moment I am using an ai manager bp, pass the EQs data to this bp and store the location in an array and check if the move to was successful. If so, then I block this location. It somehow works quite okay, but I am looking for a smarter way to solve this..

[ruby shard]

why such things happens when my character is effecting nav mesh, I want my capsule to effect navmesh (obstacle modifier)

its not always like that sometime its fine sometime no modifier at all

im using root capsule and dynamic nav mesh

normally I expect it to be like this

[uneven cloud]

You should not be using nav mesh modifiers to do obstacle avoidance. First it's expensive and second it will cause a lot of bugs with paths being invalidated.

[ruby shard]

@uneven cloud I enable it only when they are unpossessed, I fixed the problem btw, it was weird , I hade several pawn (horse, cart, ...)
sphere and capsule was not recognized by navmesh but box collision is ok

anyway it was a little tricky but it worked fine now

btw I guess it only updates if the pawn position changes, so performance is ok

[winged pasture]

Anything im missing in terms of nav mesh that could make it not work in a packaged build? Works fine in editor

( Using load level instance by blueprint, before it worked with just open level )

[terse panther]

You can do one thing...
Run your eqs in a bp and then from that take all the points and store them...and then assign those points one by one to your AI, like 1st point to 1st AI, and 2nd to 2nd and so on...

[strong cargo]

I'm trying out the EQSTestingPawn right now and finding it pretty limited since it doesn't seem to have a way to set blackboard variables? I'm working with a set of EnvQueries that grab values from the blackboard via query contexts and it looks like the EQS test pawn isn't set up to easily test more complex queries like this. Like for example, I'm using a circle generator around a target actor from the blackboard and doing dot tests that are made relative to the querier. Is there a way to use the EQSTestingPawn with this, or is it the wrong tool for the job?

At first I was thinking "maybe I can do it at runtime by possessing it so it has a blackboard/behaviortree instance" but it hides the Pawn property category so I can't set an AIControllerClass, which to me suggests Epic doesn't want people to use it in that way. So now I'm wondering if the EnvQuery itself shouldn't be using contexts like this and should instead be doing something else to make them easier to test.

Or am I supposed to subclass the EQSTestPawn and do something special to make contexts work? It seems odd to me that they wouldn't add some public properties to set up a list of blackboard key selectors/values for it to test queries with.

[misty wharf]

I think if you subclass it you could make it work yes

F.ex. in your subclass you could have it fill the BB fields or do something else to setup things for it

[uneven cloud]

The EQS testing pawn is for testing an env query in the editor not runtime. Pawns don't have controllers at editor time, even if it wasn't hidden.

If you want runtime debugging, it's far better to use the visual logger.

[rare hemlock]

@terse panther Thanks, yes this is how I am doing it at the moment. Still very unstable results yet .

[ocean wren]

evening all

Whats new?

[celest python]

GPT4

[ocean wren]

That's so old school, us cool kids are all about the gpt5 now

I do recommend playing with the gpt api btw.. very interesting to see how it functions as a model

basically you "program" it with a prompt, which is basically just a list of string pairs

super easy to do, but feels really powerful. I'm going to demo to some arts people soon.. once I've got a nice GUI for them to use

[cold trout]

i usally have a fallback in all my eqs contexts that looks for an actor tagged 'Debug' instead of using a null TargetActor, then you can visualize at editor time fairly easily

[ocean wren]

Also got a super idea for a voice based scene generator too.. but that's for next week

[harsh storm]

@celest python Quick, give me something to tell ChatGPT4 to write for AI

[celest python]

An essay about why and how using AIControlle3rs bad

[harsh storm]

Using AIControllers is not inherently bad in Unreal Engine. AIControllers are designed for controlling AI characters (also known as Pawns) in the game. They are part of the engine's built-in artificial intelligence framework and play a crucial role in handling AI behavior, navigation, and decision-making.

However, there are some potential pitfalls or misconceptions that can lead to issues when using AIControllers:

    Performance concerns: The main reason for concern when using AIControllers is that they may affect performance if not used optimally. AIControllers can become resource-intensive, especially when there are numerous AI characters in the scene or when AI logic is complex. It is important to optimize your AIControllers and their related systems (such as behavior trees, blackboards, and perception components) to minimize performance impact.

    Inappropriate use cases: AIControllers are intended for AI characters only. If you are using AIControllers for non-AI purposes or for controlling player characters, you are likely not using the right tool for the job. For player characters, use PlayerControllers instead.

    Overreliance on AIControllers: Some developers may be tempted to put all the AI logic directly into AIControllers. However, it's often better to separate and modularize your AI logic using components like behavior trees, blackboards, and custom AI components. This way, you can create more reusable and maintainable code.

    Misunderstanding the engine's AI framework: Unreal Engine has a robust AI framework that includes AIControllers, behavior trees, blackboards, perception components, and the navigation system. If you don't have a proper understanding of these systems and their interactions, you may face difficulties when using AIControllers effectively.

In summary, AIControllers themselves are not bad in Unreal Engine. They are an essential part of the engine's AI framework. The potential issues arise from misuse or misunderstanding of the AI framework, or from performance concerns related to AIControllers and their associated systems. By learning more about Unreal Engine's AI framework and following best practices, you can use AIControllers effectively in your game.

[celest python]

That wasnt a prompt that shows any potantial of gpt4 compared to 3.5 😕

I guess I am not that creative enough

[harsh storm]

I don't think it's that great of a prompt personally 😅

I've had it create the axe throwing mechanic from GoW on the first try, in Godot.
Create Doom's ticketing system in UE as well.

Also had it write a relay server in C# and remove audio tracks from an mp4 video

[dark sequoia]

I'm playing around with using perception, team affiliation and IGenericTeamAgentInterface. AIController implements IGenericTeamAgentInterface so it can perceive other actors and determine attitude. And then you put an AIPerceptionStimulusSourceComponent on actors.

The part I'm confused about is this results in both controllers and actors implementing IGenericTeamAgentInterface and providing API for GetGenericTeamId(). Does the TeamID live on the actor or the controller? Is the standard solution to make the controller's GetGenericTeamId() call its pawn's GetGenericTeamId()?

[ocean wren]

Dooms ticketing system?

[harsh storm]

It might have a different name, but this is just where I know it from. Where the AI has to queue up to do something to the player pretty much. To avoid having like 9870394857 AI attacking at once. Instead, only one or two at a time would.

[ocean wren]

Oh, gotcha

like an attack token 🙂

[harsh storm]

Yeah

[ocean wren]

I was trying to use bing chat yesterday.. what a complete mess.. only works with Edge? wtf 🙂

[harsh storm]

Yeah, dead in the water based on that decision 😅

[ocean wren]

for some reason it seems to have forgotten my login or something, so it didn't even work

I do hope that microsoft do GPT4 etc on Azure.. need something to spend my credits on 🙂

I've been trying to get DeepGram to sponsor some research so I can use their speech to text without having to pay 🙂 I even offered to make them a plugin for the UE marketplace, haven't heard back yet though

I suspect I'll end up paying for OpenAI, DeepGram and Amazon (Polly and others), which is a pain..

[celest python]

show me a code snippet for random forest algo used to implement a behavior tree in C++
this should be tricky enough

[harsh storm]

You need to ping me ya' goofball 😅

[harsh storm]

Do you want it in raw C++ or UE?

[celest python]

doesnt matter i guess, it should be able to do better logic with raw cpp

[harsh storm]

#include <iostream>
#include "RandomForestModel.cpp"

class BehaviorNode {
public:
    virtual bool execute() = 0;
};

class DecisionNode : public BehaviorNode {
private:
    RandomForestClassifier forest;
    BehaviorNode *trueNode, *falseNode;

public:
    DecisionNode(BehaviorNode *trueNode, BehaviorNode *falseNode)
        : trueNode(trueNode), falseNode(falseNode) {
        forest = RandomForestClassifier();
    }

    bool execute() override {
        double features[] = {/* your feature values here */};
        int decision = forest.predict(features);
        return decision ? trueNode->execute() : falseNode->execute();
    }
};

class ActionNode : public BehaviorNode {
public:
    bool execute() override {
        // Implement your specific action here.
        return true;
    }
};

int main() {
    // Create your behavior tree with ActionNode and DecisionNode instances.
    ActionNode action1, action2;
    DecisionNode decision(&action1, &action2);
    
    // Execute the behavior tree.
    bool result = decision.execute();
    std::cout << "Behavior tree execution result: " << result << std::endl;

    return 0;
}

it did recommend training the model in python though, using one of the popular libraries

[wheat geyser]

Hello, my nav mesh looks fine in editor but is shifted in game
No matter if I rebuild it
Any idea to refresh it correctly ?

figured it out...

I needed to delete the "RecastNavMesh-Default" that creates automatically after the first build

And then rebuild to refresh

[dawn schooner]

nice, so you are finding it useful?
I think the better you know what you want, the easier it gets. But it doesn't do a great job when asked too broad tasks

[harsh storm]

GPT4 has been quite intriguing to be completely honest. It's even implemented some stuff the way that I would've. And it is definitely much better when you really hone in on what you want.

I definitely was not impressed with the previous version.

But this one...dear lordy.

[dawn schooner]

are you paying or is it the free version?

[harsh storm]

It ain't as great as modern plumbing mind you. But it is still pretty darn nice.

I'm paying.

[dawn schooner]

👌

[harsh storm]

I made a deal with my friend. If it could implement that GoW mechanic, I'd sign up right then and there.

[dawn schooner]

yeah I get it

[harsh storm]

And I did.

[dawn schooner]

hahaaha that is a good test

[ocean wren]

Isn't it currently gated?

I'm already paying for the old one 🙂

[harsh storm]

The API is gated I believe, but the prompt one you do on the site, as long as you're paying that monthly fee, you should have access to it.

[harsh storm]

It was impressive to me that it knew exactly what I was talking about

[ocean wren]

oh gotcha.. yeah

gpt+ or whatever it is

[harsh storm]

At first, it didn't implement the arc throw, so I told it to include the arc. Then it rewrote it to include a variable for the arc.

Worked without me having to change anything

[ocean wren]

The history is really the key to it

[dawn schooner]

impressive

[ocean wren]

And this new model has up to 32k tokens, which is bonkers

[dawn schooner]

ye most times it does not work for me, I know I'm being too vague or imprecise

[ocean wren]

I'm hoping for it to generate me a monologue (assuming I don't run out of cloud credits on deepgram or amazon 🙂

[harsh storm]

With this version, I think it can be a force multiplier with someone who understands the underlying system.

[dawn schooner]

also I found it is not very good at answering "how to" questions

[harsh storm]

I definitely wouldn't trust code blindly.

[ocean wren]

We're doing a research project based on movement as prompt.. hoping to get some behavioural movement in these multi-modal models

I think chatGPT and the like is actually an ideal use case for a "coding buddy", where you need someone to remind you of syntax and the like

a bit like an AI pair programmer 🙂

[harsh storm]

I even asked it to write a game idea about mushrooms.

Now my wife wants me to make the game 😅

[dawn schooner]

LOL

[harsh storm]

But yeah, it's going to be a great buddy programmer...for people who can tell when somethin' ain't right.

[ocean wren]

I use it for python, because I can never remember the syntax, but know what I'm doing otherwise 😉

[dawn schooner]

I would find interesting to see if it's capable of generating new solutions to existing problems or to solve new problems

[dawn schooner]

xD that is me writing Python too

[ocean wren]

Its causing quite a few issues at the University though.. because obviously students are gonna use it

[dawn schooner]

ye

[ocean wren]

Had a presentation on AI from someone in health and social science last week 🙂 was pretty fun to see someone else not in CS get whats going on 🙂

Its going to transform education pretty widely

[dawn schooner]

indeed

And not necessarily for the worse

[ocean wren]

Scary how few people at the upper levels even know it exists though

Yeah, personally I'm happy for the current system to die

[dawn schooner]

+1

although I understand it can be concerning in the short term for certain situations

[ocean wren]

Its about time the industry of education had a shake up

[harsh storm]

Seriously. It is so archaic

[ocean wren]

It just doesn't bloody work, I mean lectures are the dumbest thing ever, but thats what we do.. because thats what we do

[dawn schooner]

ye

[ocean wren]

And they just keep increasing class sizes, so you've got sod all chance to help anyone

[dawn schooner]

I was watching my nephew (11) doing some homework and studying... damn it was painful to see him lose his time like that

[ocean wren]

They need to stop turning it into a money making venture

I mean almost all UK uni's are supposed to be charities, but of course they make billions and who gets it? the VC's and senior managers

[harsh storm]

Ain't much better in the states

[ocean wren]

States is mucho screwed 🙂

[dawn schooner]

in Spain we have public unis, which at least means you don't get in debt to study

but quality varies widely and it's still very archaic

[ocean wren]

Yeah, we used to have that.. but then they introduced fees

[dawn schooner]

and it's still not cheap

[ocean wren]

I just wish they made it so that people could study when they're ready and not just when they leave school

[dawn schooner]

but the education issues (at least here) come from the bottom

[dawn schooner]

indeed

[ocean wren]

I'd rather avoid kids who just aren't interested honestly

[dawn schooner]

exactly

[ocean wren]

Doesn't do them or me any good 🙂

and my dreams of a ninja game dev army aren't coming true anytime soon

[dawn schooner]

putting kids 6-8 hours per day to learn things they are not curious about in the first place is a recipe for disaster

[ocean wren]

Yeah

impossible

[white pasture]

what's a way to do delayed sight perception in c++? IE where the sight stimulus has to be active for a time before it triggers the AI noticing you. Something like in this video:
https://youtu.be/O5o_usc2pWo?t=461
maybe something where there's a float value that goes up while you are in sight, and down while you are out of it, and triggers awareness at a certain threshold? but in c++ (video is in BP's, and i'm not sure of c++ equivalent to On Target Perception Updated). thanks!

[lyric flint]

Override the On Perception events in the AI controller and add your threshold logic there is one way

[ocean wren]

I built mine on top of the perception system and it worked exactly like that. Mine had a bunch of different factors in though, things like concealment, perception curves for different ranges and angles (percieve more if something is in front of you rather than in the peripheral vision) etc. Essentially you take the output of the perception system and feed it into another system that does the actual perception scoring.

[haughty coral]

https://github.com/EpicGames/UnrealEngine/commits/ue5-main?after=b871db910e849524b9a88345a946493f0cb8a77e+69&branch=ue5-main&qualified_name=refs%2Fheads%2Fue5-main

[lapis fractal]

You fixed it?

[pearl fern]

@lapis fractal No. Just a dirty workaround of setting the navmesh from static to runtime generation. That makes it work in a packaged build, but it's obviously not ideal.

[lapis fractal]

Extremely strange behavior, considering that I disabled WorldPartition

[lapis fractal]

But thank you at least for this decision)

[pearl fern]

@lapis fractal . No problem. Let me know if you ever find a real fix please ^^

[celest python]

nice

[lapis fractal]

Ok)

[ornate sleet]

guys is it normal that the behavior graph of my AI are like this when running ? you see the boxes are blinking

[harsh storm]

It is happening like that because the tree is continously reevaluating the tree. The task most likely is ending before the animation is actually finished.

[ocean wren]

Yes its normal, your "has reached character" value is true.. so it selects attackplayer to run because the condition on your chasing is not allowing that to take priority

Normally, you'd have an idle with a wait to take up idle time

Your "idle" time is basically looping attack player, which seems wrong

[ornate sleet]

Okk

because I can't figure out why after like 30 sec my game start to freeze each 5 seconds

[ocean wren]

use unreal insights to capture the performance and find out

[ornate sleet]

Ok I can see the look of the engine is really long sometimes

How Can I know what is causing this ?

What is this wait for task who took 500 ms wtf

[ocean wren]

You need to watch some videos on using insights

[ornate sleet]

@ocean wren I watched some, but can't really go further, I mean I see it's the loop tick and wait for task which are too high

And the problem doesn't appear when I package the project

[white pasture]

Thanks! And you did that with new aisense or senseconfig children? How do you access on update or tick with those?

[strong cargo]

Yeah, I wasn't interested in using it at runtime. I was asking what the "right way" to use the EQS testing pawn. I did something akin to what @cold trout suggested. A little messy, but in the interest of time I made a component with context-related variables I cared about, attached it to the testing pawn, and used that as a fallback if available when a controller/blackboard couldn't be found in the query context. When I have more time I think I'll extend the testing pawn to take a blackboard asset and an array of key selectors/values to simulate a blackboard that can be swapped out in contexts when a controller/blackboard can't be found.

Afterward I saw there was a way to bind query params to actor properties (I think? Haven't tried yet so maybe I misinterpreted the engine source code) which would mean subclassing the testing pawn probably would've been the better move, but I was pretty intent on using the blackboard.

[ocean wren]

You need to drill down to find the issue. Those are tasks which contain other tasks right? so they're capturing the overall task time and not clearly identifying the issue.

[ocean wren]

I had a few custom AI senses yeah. But for the most part I just built a wrapper around the perception system as was and simply read the perceived senses and then did the scoring part as a second pass. So my own systems relied on the secondary system (i.e. it didn't percieve directly from the perception system, but rather the wrapper). The wrapper took its updates from the perception system and did the scoring per tick.

[white pasture]

Thanks!

[ocean wren]

I also built a debugger, which was pretty important for testing.. essentially it showed a graph of perception for whatever I was pointing the camera at (if I recall), so you can see the values going up and down and also visualize the threshold for when something is "interesting" vs "percieved" etc

You'd maybe use it for UI stuff if you're showing the perception to the player

But our use case was a bit more hardcore, so no UI, but it was great to use for debug anyway

Here's what it was like: https://www.youtube.com/watch?v=Qo-6AEyda98

[white pasture]

Awesome! I assume this project is not open source?

[ocean wren]

Nope, they got published by Microprose

I was mostly just helping them out for fun and to do something interesting

Oh yeah, I forgot.. I also used ADSR envelopes (attack, decay, sustain, release) for the different senses

[white pasture]

Like decay in intensity over time after initial sound report?

[ocean wren]

exactly

its basically the same stuff you get in audio, but for perception shaping

means you can fiddle with values in a way that makes sense to designers

[white pasture]

Damn I am not on your level lol.

[ocean wren]

I was mainly just playing around 🙂 but it definitely worked

[ornate sleet]

@ocean wren yes but how I can the details of these tasks ? Like the engine loop one for example. And I don't have the problem when I package the project maybe it's some debug stuff which run in background and sometimes make it freeze ? Is it possible ?

[harsh storm]

Ehh - zoomba has a fetish for writing tools. Don't sweat it 😅

[ocean wren]

You can drill down into the task graph and see what its made up of, try looking at some of the sessions on using the profiler.. it really does pay to learn to use it.. And yeah, things like the garbage collector could be causing some kind of stutter.. only way to know is to profile

[ocean wren]

Hahaha.. too true

You got me there

I still think that making tools is a most valuable thing for any game production

I've just setup a new company to make AI powered tools for games 😉 so I guess I would say that

[harsh storm]

It is. At least according to old school idTech. It was one of their principles, according to Romero.

If he had 30 days to make a game, he'd spend 20 on making the tools.

[ornate sleet]

@ocean wren you really can see into the FEngineloop:tick ??

[ocean wren]

They had one of the best examples really.. making FPS's really worked to make the genre more productive.. DoomEd and the like

WorldCraft, whatever the Duke Nukem editor was called.. Duke Ed? 🙂

[ocean wren]

You can profile all of the parts of a given tick yes

[ornate sleet]

@ocean wren any doc to know how do that ?

@ocean wren only using the insight ?

[ocean wren]

insights is the way to do it

epic have done some livestreams on it, but I think its on the learn site too

[white pasture]

Anything released yet lol?

[ocean wren]

No, only just setup the company.. probably be doing some consulting gigs initially

But we've got a bunch of research projects that I think will eventually lead to actual products

No idea.. maybe there's a linux channel you can ask on?

ah, yes, its there in PLATFORMS on discord

I would assume so though

Man, Unreal Engine is so random.. for some reason the water in one level just.. stopped working 🙂

Mind you, I did go from 5.0 to 5.1 so I guess they killed it

[pale schooner]

Is there a compile button in the behavior tree?

[sacred shale]

Is there a way for a static mesh actor to ignore creating navigation data on top of it but still be part of the navmesh (as in, creating boundaries around it)?

I'm running into a couple issues where I have really small box static mesh actors that generate navmesh data on top of them, so my AI walks on top of the crates and gets stuck (and yes I have turned off Can Character Step Onto)

[ocean wren]

There's a navmesh blocker you could use

[sacred shale]

How would I add that? Just googled it and I can't seem to find it

Hm yeah it doesn't look like it's what I want

I guess it's impossible it's probably the way my capsule is pushing against small objects

It shouldn't be stepping up on boxes in the first place :/

[ocean wren]

look at the step height?

[sacred shale]

step height doesn't do anything, but walkable floor angle does

but I don't want to change the default walkable floor angle

since I've run into instances where it will start falling down a minor slope :/

[pale schooner]

Is there a specific reason why MoveDirectlyToward works and MoveTo does not when given the same object in behavior trees? (Common mistake)

[pale schooner]

More info, MoveTo is firing but not moving, but if I replaced it with MoveDirectlyTowards it would work

[celest python]

check visual logger

[pale schooner]

Don't know what this proves or shows tbh

[celest python]

click to the keyframes

see logs

it will explain why its not working as you intended

or at least give a hint

[uneven cloud]

No. It's a data asset, not a BP.

[uneven cloud]

Lack of a path. Move to uses navigation to move along a path. Move Directly Towards does not.

[sacred shale]

could the capsule have anything to do with the fact that my AI is walking on top of objects that explicitly have the "Can Character Step Up On: No" property???

I swear I've spent hours on this and I've found two solutions: 1. is to reduce my capsule radius to not be like a sphere (but that messes with the other logic in my game) 2. reducing the walkable slope angle (which messes with other logic in my game) I'm seriously lost, how have few people ever run into this?

I have looked at at least 10 different forums and it seems more like an Unreal bug specifically with spherical capsules

[floral brook]

Hey there guys! First time even touching the navigation stuff that exists in UE - is it possible to using the navmesh for pathfinding without hooking up an AI to it? Like passing the start and end and getting out a path?

[dawn schooner]

yeah, it is (although you'll need to potentially handle corner cases yourself)

I can look into the computer tomorrow, I do it for battalion-based pthfinding

[floral brook]

Hell yeah, I would really really appreciate that

[dawn schooner]

this is my code for pathfinding, take into account it is probably not very refined yet, also there are lots of hardcoded values

UNavigationSystemV1* NavSystem = FNavigationSystem::GetCurrent<UNavigationSystemV1>(GetWorld());
if (!IsValid(NavSystem)) return;

Path.Empty();

FNavAgentProperties NavAgentProperties;
NavAgentProperties.AgentHeight = 100;
NavAgentProperties.AgentRadius = BoidConfig->Radius;
NavAgentProperties.bCanWalk = true;
NavAgentProperties.bCanFly = false;

FPathFindingQuery NavParams;
NavParams.EndLocation = FVector(TargetLocation, 0);
NavParams.StartLocation = FVector(ReferencePoint, 0)
ANavigationData* NavData = NavSystem->MainNavData;
NavParams.NavData = NavData;
NavParams.QueryFilter = UNavigationQueryFilter::GetQueryFilter<UNavigationQueryFilter>(*NavData);

FNavPathQueryDelegate PathfindingDelegate;
PathfindingDelegate.BindUObject(this, &URTSBattalion::OnPathReceived);

NavSystem->FindPathAsync(NavAgentProperties, NavParams, PathfindingDelegate, EPathFindingMode::Regular);

[ocean wren]

I saw a video on twitter the other day, apparently someone did a new navmesh generation so you can update the navmesh in realtime in such a way that you can like hand edit it (the CSG was so much faster)

[celest python]

You mean glassbeaver's thing?

[ocean wren]

Might be? 🙂

just saw it randomly on twitter 🙂

I like seeing new stuff like that

feels like progress

[celest python]

GlassBeaver saying recast couldnt be optimized further without unreadable assembly logic and stripping some logic proves Mikko did a great job on it 😄

[ocean wren]

Yeah, but almost everything can be improved.. new ways of thinking etc

Not something you see a lot of though, people tackling the harder stuff you know.. its mostly "why doesn't X move" 🙂

Which suggests that the UX for the current tools is... not great

[dawn schooner]

ye but that is probably not Recast's fault, more UE's wrapper

[ocean wren]

Oh its definitely not

[celest python]

"why doesn't X move"
its BT editor's fault

[dawn schooner]

I think recast already does dynamic partial updates

[ocean wren]

What I mean is that I like seeing people improving tools and trying new ideas... I guess I learn more by seeing things working.. "build it to understand it" approach

it does real-time CSG yeah

I think the BT editor could definitely use a makeover.. making failure cases more obvious (and the cause of failure)

[celest python]

zoom i have a question
how much costly it would be to run a gpt model to stream something on twich, assuming i have a cool tweaked gpt model in a server

i heard it requires 40 GPUs to run

[ocean wren]

It will use a kubernetes cluster most likely, at inference time who knows.. probably not THAT bad

there's some open source efforts to create similar models ongoing (Bloom etc)

they need a ton of GPU for training, but for inference usually a lot less

Why would you want to run your own server though? API not enough?

I mean I want to run stuff locally too, but the API is usable

[celest python]

I guess it is, but with long/short term memory thing it consumes a lot of tokens

[ocean wren]

new GPT4 will have 32k token length max... which is a crazy amount

[celest python]

GPT3.5 consumes 1k tokens with short term memory per a few sentences i guess

which means $0.002 per response

[ocean wren]

I found out that Azure has the new chatgpt model available, so technically I can use my Azure credits for generation, but I've already paid some creditrs for OpenAI api access

you actually get the number of tokens used for both request and response when you recieve the response, there's a python package to estimate pre-call too

I don't think 0.002 per 1k tokens is so bad.. GPT3 was a lot worse

[celest python]

did you ever implement memory context to any GPT?

[ocean wren]

And I'm sure open source models will be around very soon that do something close

yeah, its part of the chatgpt api

[celest python]

(sorry im still quite noob yet :D)

[celest python]

what do you think about the idea having another cheaper model to summarize the latest convo to a more advanced model to implement it

would it work

[ocean wren]

you basically give it a list of text pairs.. the first text is just identifying the context of the second.. so you can use "system" to mean that the next text is a system message

Yeah, chaining models definitely works

I've got plans to have two gpt's, one for dialogue and another for command and control of the environment

the "memory" is just a bunch of text too 🙂

[grave compass]

Which one should I use "ai perception" or "pawn sensing" ?

[ocean wren]

in fact, the whole thing is just a bunch of tokens

AI Perception, Pawn Sensing is deprecated

[celest python]

If I can, I'll try to do something like this when I have more free time https://youtu.be/r92V7WMPTtg

its basically a GPT reading chat history and developing a memory context from it

and replies to random questions

this is actually similar to your metahuman gpt thing 😄

[ocean wren]

Yeah, there's a load of fun stuff to play with

its strangely compelling

We are still waiting for our mocap hardware to be installed, once that's done, I'll be capturing a TON of mocap for use in the metahuman context

I'm planning on making a model that does the gesture of speech as part of the viseme generation from voice synthesis

so basically, you get body language from spoken language

which is one area that sucks with the amazon polly approach I've got now

polly generates facial visemes.. but the body is disconnected still

[celest python]

oh my god can we make GPTs rap synchoronously with a beat?? https://youtu.be/Zp37voNDY_0?t=227

damn AIs make me feel like I'm from stone ages

[ocean wren]

There are AI singers you know

[celest python]

[ocean wren]

Helps if you use more of a synthetic style.. but I reckon all AI music generation will become the norm, especially for games

And within a few years

[celest python]

I think 3D art part will be more difficult, I've been brainstorming about providing input to an AI from level data to make it optimize lights and place props accordingly but it all happens in a 2d perspective

couldnt find a way to collect proper data for it based on static mesh data and such so it will both arrange a scene and handle collisions well

[ocean wren]

I think the main issue for the 3D stuff is compute.. I suspect that we'll use NERF's to generate the basis for it

I think you're thinking about this the wrong way.. we'll generate the scene from an input image, give it an image of the scene you want and the engine will create it

OR we'll use voice to describe it

[celest python]

but meshes and light entites/actors exist in 3d matrix world

it would have to project from 2d

[ocean wren]

"I want a large terrain like mount fuji, with a lush jungle at the foot of the mountain"

naah, it exists as concepts in the latent space.. the scene could literally be described as a graph from that latent space

no hand editing required basically

[celest python]

I dont get how AI will be able to properly setup back of the mount fuji for example, its only aware of the 2d representation of it from camera i guess?

[grave compass]

I see that they still use pawn sensing in newly shared tutorial videos. Why are they using it?

[celest python]

I guess maybe it can work on something like a UV map right?

[ocean wren]

naaah, forget what we've got now.. imagine you've trained on satellite data from the world

blackshark.ai already do that

[celest python]

dont follow those tutorials

[ocean wren]

google has scan data of most of the world

[grave compass]

ok so i will use ai perception

[celest python]

Imagine that AI exposes some hidden military bases by mistake xD

blackshark.ai is amazing

can i find some technical explanation about it in web?

is there something like GDC but for ML?

[ocean wren]

Epic did a livestream with them a few months back? should be on their channel

[celest python]

so its UE based?

even better

[ocean wren]

Well, blackshark is used in MS flight sim.. but they I think got bought by Maxxon, who do satellite data from their own sat network

And there's a plugin on the marketplace I think..

GDC for ML is probably ICCV, CVPR or NeurIPS 😉 all those are huge ML conferences

some of the most fun ML stuff is presented at Siggraph though, which is the main graphics conf

[scarlet oriole]

Hey
I was working on our behavior trees and was wondering how to limit the max number of enemies to attack the player at the same time?
The naive solution is to have a blackboard integer that increments when an enemy attacks, But this comes with the problem of decrementing both when the enemy is destroyed and when it stops attacking
Same issue if I have some array
There must be a better solution, what might it be?

[harsh storm]

Ticketing system.

You can look up how Doom did it

A simple example would be:

1. World Subsystem - this has operations that allow you to request a ticket, release a ticket. As well as the max number of tickets available to be given at one time. (Use UDeveloperSettings to make it easily changeable in the editor)

2. When the AI wants to attack, request a ticket - if it can't get one, place it in some kind of background task to have them waiting for one. Put them in a queue pretty much and while waiting for a ticket, make 'em appear to do something. Like they're winded or something, idk.

[celest python]

https://github.com/intaxwashere/AITokenSystemUE

[scarlet oriole]

Thanks guys! Yeah that system seems a lot smarter will definitely do a dive into this

[celest python]

depending on your game design you might need a top-down system rather than bottom-up one like DOOM's

spiderman GDC i mentioned in the repo demonstrates a topdown one

i.e. manager decides tokens

[celest python]

systemic AIs = bottom up
authored AIs = top down

[vale notch]

Hello everyone. I'm having serious issues with nav queries. I have 3 nav areas, clearly visible with their respective colours when viewing the nav mesh, and I have one nav query filter that has a travel cost of 1 in the center area, and a travel cost of 1000 for the left and right areas, yet my AI still prefers taking the left route if it's open, otherwise it'll take the right. The actual distances are pretty similar, maybe 20 meters or so for either path.
I am simply using this AI MoveTo, and the filter class is set to the correct nav query filter. Any ideas why it's taking the more expensive route(s)?

[vale notch]

The default costs for the areas is reversed, so if it is using the default costs the behaviour makes sense. I'm not sure why they don't seem to be using the override values.

[vale notch]

So it turns out the GetRandomReachablePointInRadius node doesn't apply the nav query filter to the whole route, meaning it's using the default travel costs.
Now my problem is that this node, which does work as I intended, doesn't have an "On Success", so I'll have no idea when the AI reaches the destination. Anyone know how to figure that out, or if there's another node that would work that does have an "On Success"?

[misty wharf]

not 100% sure but you might be able to provide a default nav query filter on your pawn you're moving using the AITask version

I don't remember where it was, maybe it's on the AI controller class where you can specify it

iirc AITask_MoveTo should use the one specified as the default, which might be why you're not getting the result you want

[vale notch]

That is correct. The Move to Location node works though as it does get the Filter Class option, and I realized I can use this bind to move completed to know when it has arrived. Thanks though! 🙂

[misty wharf]

Nice :)

[modern owl]

how do I get all actors of a certain type within an eqs query? I want to generate a grid around the player and get all pawns of type myPawn in that grid

[blissful bridge]

hi guys, i have a problem with pathfinding, when they start moving with move to they follow this strange path, i try every possible combination for the generation of the navmesh but nothing change

this is the generation settings, heurystic scale = 0.999

[solid echo]

Heyo, having an issue with navigation on this bridge object, the ai doesn't know to go onto the bridge because the path is above instead of below. I made a new Nav System in project settings, no luck. Any help would be appreciated!

[vale notch]

You should look into your collision models. The nav path shouldn't be on top of it like that unless its collision is not allowing them to be on the bridge.

[solid echo]

Thought of that too, I made 3 simple box collisions to mark the boundaries

Here's the collision settings for ref

[vale notch]

That "nav system" you made though is not used for what you seem to be trying to use it for. You added a nav agent, that's the character/pawn and its size, it has nothing to do with your bridge.
The collision does look alright, so it's a little strange that it generates the nav mesh like that. 🤔 What does your other nav agent look like?

[solid echo]

Nothing special, I made a copy of the default for reference

You get the same result but in lime green lol

[vale notch]

Hmm. Is there anything else there than the bridge mesh?

[solid echo]

This particular one has a spline that lets me drag it out to get this shape, but the static mesh has the same problem

[vale notch]

Okay. Well it has to have something to do with the collision. You could try dragging the simple collision boxes for the railings down to the same level as the rest of the bridge, just to check if that changes the nav mesh generation.

[solid echo]

Even if I delete the ones on the sides the nav stays on top

I turned on view complex collisions in case there was a hidden collision but there doesn't appear to be any

[vale notch]

Quite fascinating. 🤔 I'm not sure what the problem might be, but it's clearly using some form of collision that encompasses the whole thing. You can try searching for "Collision Complexity" in the details panel and changing that to one of the other options and see if that changes anything.

[solid echo]

Oh I figured it out so dumb

It was the collision on the bottom, there's a diagonal line going across, that was the complex collision

I replaced it with a simple box and it works now 🙄🤣 Thanks for your time, much appreciated

[vale notch]

Alright then! 😄 Glad you figured it out.

[sacred shale]

is there a way to increase vertcial sight perception? I'm always too high or too low above my enemy and I want to make sure that it never loses sight of me regardless of the height

[uneven cloud]

Using an EQS context. You need to make one to get the actors that you need.

[uneven cloud]

You'd need to make your own sight sense.

[modern owl]

a context on what test?

[uneven cloud]

That depends on what you are trying to do.

[sudden citrus]

UBlackboardKeyType_GameplayTag should really live in the AI module or somewhere more non-committal than GameplayBehaviors

before someone says "put a PR in", how would this be handled in terms of backwards compatibility?

[modern owl]

I just need to get a reference to the character class of all of the characters in the grid

[sudden citrus]

I think there's some method like "can actors see each other" to override on the controller

[final loom]

How do I set the AI teams as part of the stimuli component?

[dark kiln]

Can someone tell me how to interrupt a behavior tree sequence if the player is sensed? I have a blackboard decorator trigger on sight of the player, but it's not firing off when I walk in the path of the enemy

[final loom]

Faithdrawn; do you have it the AI perception component detects neutrals and friendlies? I was having that issue too.

[dark kiln]

Yes

I checked all of those boxes

[final loom]

Ah okay. That was the immediate thing to check for me, do you have some kind of debug message to test that the behavior tree's blackboard is getting told it sees the player?

It may not getting assigned that, could be a factor.

[dark kiln]

I don't but that's working. I have the enemy running to specific points and only if I'm standing in front of him before he moves to the next one does that selector get called for him to chase me. Essentially I need to interrupt the wandering to chase the player if I'm seen

[final loom]

Alright. Could maybe be an order of operations error

[dark kiln]

Perhaps. I have wander on the left and chase on the right. I am going to set the chase to only occur when you agree to be in a fight (it's a human NPC you fight) so maybe I shouldn't worry about it right now unless it so doesn't work when I restrict the chase to the combat

So if I want to call the black board decorator outside of the BTtask, do I set up a variable for a behavior tree or for the key selector or something else?

Basically going to set it up only to run after the dialogue event sets in combat to true from the player

[ocean wren]

If chase is higher priority, then it should be the leftmost behaviour. Priority is left to right order, higher to lower.

It should also "aborts lower" so that if it becomes able to run, will take over from the wander

[terse panther]

@celest python @uneven cloud @ocean wren a small question , like after completing the basic of the AI in unreal, like BT, EQS, perception along with custom senses and custom eqs, what would you guys recommend to learn/try next? i am getting a hard time to decide the things to learn about advance AI ,
can you guys please suggest something?

[celest python]

After some point advanced topics start to reach a point where you need to know internals of low-level systems, zoombapup and Luthage would know better about those cases

But after a point it turned to a passive learning process for me by reading the source and asking dumb questions here

[terse panther]

for now i am reading the built-in eqs generators, and getting some good knowledge from them...

[celest python]

You can research about GOAP, HTN, A*, UPathFollowingComponent source code, Ubisoft's AI GDCs, uncharted authored vs systemic AI talk and game ai pro 360

along the way you should see a lot of new keywords to research about too

[floral brook]

Excellent, thank you!

[terse panther]

Thank you for all these topics....
Yesterday I learned about A*, how we use, why we use... Now thinking to may be try to do a little code with this...

It's my first time hearing about HTN 😅

For the GDC talks i wanted a lot, but some are still missing ....

[celest python]

I'd research how AIs are designed in other engines, because UE itself doesnt have a great design at all

concepts are same or similar but tooling differs a lot, and it affects design

[harsh storm]

Before moving on, I'd recommend actually doing a small game or two using A*. So you can really cement the knowledge.

A* is a foundational algorithm in many things to be honest.

[wary ivy]

program a min heap while at it too 😛

[ocean wren]

I guess it depends on your interests.. certainly there's a lot of reading about different pathfinding problems.. crowds and the like (look at Detour and RVO in UE and improve them for instance). It would be a good idea to look at Utility systems and build one of those. Personally I've veered quite a lot into machine learning and reinforcement learning approaches for generative AI. But then I'm trying to create a Holodeck, so I doubt its something you'd want 🙂

Also, learn slate and UI and tools development

Oh and Control Rig 🙂

Oh and animation systems like blend trees and the like.. AI programmers often end up having to partly be animation guys to some extent

[celest python]

AI programmers often end up having to partly be animation guys to some extent
i hate this fact

i think control rig is quite overkill though haha

[ocean wren]

Ubisoft is a good place to look at for character stuff to work on.. they've got some excellent presentations and problems to solve

Control rig is fun!

[celest python]

It's VM is good

[ocean wren]

Doing some digital puppetry demo's with it

The research group I'm with has quite a diverse user base 🙂

Mostly involving motion and animation right now, ML/RL driven stuff

[harsh storm]

sigh - I hate that structs can have UFUNCTIONs in them in that VM but not the main VM 😅

[celest python]

its really amazing how half of the modules can be perfect and other half can be full of wtf moments in UE

[harsh storm]

I bet they know how they can fix some of this stuff but can't due to licensing reasons.

[ocean wren]

different people with different experience writing them

[harsh storm]

I wish they'd just...break stuff sometimes (in a good way)

[ocean wren]

licensing reasons?

[harsh storm]

Having to maintain backwards stuff

[ocean wren]

Oh gotcha 🙂

[celest python]

Its about team imo, they rewrote RigVM, which means they were really not half assing for benefit of Fortnite

[harsh storm]

Or making sure things don't break too much or w/e.

[ocean wren]

I was wondering who they'd license code from 🙂

[harsh storm]

Nah - it's more for people who license the code from them 😅

IE - us

But let's be real...other AAA studios

[ocean wren]

RigVM is more for their studio use, so doesn't have the same delivery schedule maybe

[celest python]

I guess so, but in general after some version of UE animation team did great stuff without any UX issues

thats something unusual for UE 😄

[ocean wren]

Yeah, maybe they had their UX team actually work with them on that

Its hard to tell because some parts of the engine has terrible UX, other parts are pretty good

maybe a coin flip?? 🙂

[celest python]

Did you see Verse?

[ocean wren]

that the new scripting language?

[celest python]

yeah

[celest python]

If you have any idea what line 141 is doing, share with us 😄

[ocean wren]

That looks like some kind of assignment? so maybe its an optional assignment?